Speech Signal Based Broad Phoneme Classification and Search Space Reduction for Spoken Term Detection
G Deekshitha, Mary Leena · 2018
Depending on the acoustic properties, sound units can be grouped into six broad categories as Vowels (V), Nasals (N), Fricatives (F), Approximants (A), Plosives (P) and Silence (S). This paper proposes a set of signal based features for broad phoneme classification. A Deep Neural Network (DNN) based Broad Phoneme Classifier (BPC) is trained to predict the broad phone class labels in continuous speech. Across the TIMIT test data, the BPC gives an overall frame level accuracy of 75%. After smoothing the BPC labels, 83% of the broad phone regions are correctly detected by the system on an average. These broad class labels are useful for applications like speech recognition, and Spoken Term Detection (STD). Effectiveness of the proposed broad phoneme classifier labels for search space reduction in STD is illustrated using TIMIT database. Around 82% of search space reduction can be achieved for detecting 78% of the keyword locations from the TIMIT test dataset.